2014 7th International Conference on Intelligent Computation Technology and Automation

An intelligent E-commerce recommendation algorithm based on collaborative

filtering technology

Yang Xiao qing

Henan Technical College of Construction , Zhengzhou City, Henan Province 450007, China
[email protected]

Abstract—This paper presents an intelligent E-commerce that is, personalized recommendation. Recommendation
recommendation algorithm with collaborative filtering systems utilize knowledge discovery techniques into
algorithm. Firstly, a novel user interest model is given, which personalized recommendations for products in the process of
is an important module in the E-commerce recommendation online transactions[4]. Recommender systems are used by E-
system. Particularly, to effectively integrate the user interest commerce Websites to suggest products to their customers
model with collaborative filtering algorithm, we assume that if and then to provide consumers with useful information to
two users have similar interest vector, they may want to choose assist them determine which one to buy. Particularly, there
the same products. Furthermore, we define the users with
are many difficulties in the recommender systems, one of
similar interests as neighbor, and finding the neighbors is of
which refers to the ability to be adaptive to environment
great importance in the E-commerce recommendation.
Secondly, the intelligent E-commerce recommendation
where users may have many different preferences or
algorithm is proposed based on user-rating matrix, and the products have diverse contents.
products with highest scores is recommendated to the target Collaborative filtering means a technology which utilizes
user. Finally, experiments are conducted to make performance. known preferences of several users to predict the unknown
Compared with other two schemes using four metrics, it can preferences of a new user and recommendations for the new
be seen that the proposed algorithm is more suitable to be used user are based on the former predictions[5][6]. However, the
in E-commerce recommendation system. traditional collaborative filtering systems can not provide
accurate recommendation, because the predicted items for a
specific user is not identical to the interests of his
neighbors[7][8].
The rest of the paper is organized as follows. User model
is introduced in section 2, which is an important part in the
Keywords- Recommendation algorithm, Collaborative proposed E-commerce recommendation system. Section 3
filtering, Weight matrix, User similarity, Neighbor illustrates the E-commerce recommendation algorithm based
on collaborative filtering. In section 4, experiments are
conducted to make performance evaluation. Finally, the
conclusions are drawn in section 5.

I. INTRODUCTION

II. USER INTEREST MODEL

As is described in Wikipedia, E-commerce represents
trading in products or services conducted through Internet.
Several technologies are integrated in the E-commerce In this section, the user interest model is given, and this
platform, such as mobile commerce, electronic funds model is an important module in the proposed system.
transfer, supply chain management, Internet marketing, Firstly, we define the interest vector for user ui as follows.
online transaction processing, electronic data interchange,
and so on. In recent years, E-commerce platform usually V
u j  
vi1 , vi 2 ,
, vim ,
, vir  (1)
utilizes the World Wide Web at least at one step in the whole
transaction process[1]. On the other hand, E-commerce is
generally regarded as the sales aspect of E-business, and it is where vim is the weight of sim , and the conditions
made up of the exchange of data to facilitate the financing r
and payment for the transactions. Therefore, E-commerce vim  [0,1] and v
m 1
im  1 are satisfied. Hence, the main
refers to an effective and efficient method to communicate in
an organization[2][3]. problem is to achieve the optimal solution
For the E-commerce platform, it is of great importance to
recommend products to users according to their preferences,

978-1-4799-6636-3/14 $31.00 © 2014 IEEE 80

DOI 10.1109/ICICTA.2014.27
V
u j   
vi1  ,
vi 2  ,
,
vim  ,
,
vir   from the user
* * * * * the problem of E-commerce recommendation can be
 described as an optimization problem as follows.
interest matrix M i .



2
In the interest matrix M i , the row Ri  ri1 , ri 2 ,
, rih
Tij 
UV T 
2
min  W (T  UV T ) (4)
ij F
means a discrete set of items and the column i , j

Ci  ci1 , ci 2 ,
, cir .
In order to combine the user interest model with Subject to: U  R md , V  R nd
collaborative filtering, we suppose that if there are two users
sharing the similar vector, they may want to buy the same where  represents the set of i, j , in which entry Ti , j is
products.
The main idea of collaborative filtering technology lies not missing, and is the Hadamard product. Based on the
in that when two users have similar ratings on specific above analysis, the proposed algorithm is given as follows.
products, they may have similar interests. Hence, in this
paper, the recommendation results may provide to users
based on their neighbors. The neighbors of a particular user Algorithm 1: Intelligent E-commerce recommendation
ua are defined as follows. algorithm based on collaborative filtering
Input: user interest vector V
ui  ,

Neighbor (ua )  ui  U sim
ua , ui  sim
ua , ui  (2) Output: top ranked K products ( p1 , p2 ,
, pk )

j[0, m ], i j
(1) Define the set of features which describe the user
interest S
ui   si1 , si 2 ,
, sir , and sir is a feature for
user V
ui  .
where sim
ua , ui  means the similarity of interest between (2) For i  1 to N ( N represents the number of users in
ua and ui . For user ui and u j is obtained based on pearson the training dataset)
(3) For j  1 to N
correlation coefficient as follows.
(4) If i j
(5) Computing the similarity between ui and u j by Eq.3.

r u j 

 ru j  rui  rui  (6) For each user pair, sorting the user similarity using the
sim
ui , u j  
 Oui u j
(3) descending order, and then select the top M ranked
users, and these user are denoted as M
ui 

r
u j  ru j  
r ui  rui 
 Oui u j  Oui u j (7) For each user in the set U  ui
(8) Computing the predicting score of user ui ’s ratings as
where Oui u j refers to the set of common products of user follows.
ui and u j , and rui , ru j represents the average of ratings of  sim(i, j )  t
k M i
ij

user ui and u j . (9) psij 

 sim(i, j )
k M i

(10) Sort all the predicting score psij with descending order
III. E-COMMERCE RECOMMENDATION ALGORITHM (11) Select the top ranked K products ( p1 , p2 ,
, pk ) to
BASED ON COLLABORATIVE FILTERING user ui

In the E-commerce recommendation systems, there are Using algorithm, the top K products with highest scores
only a few of products that user may rate, that is, the user- are recommended to users
rating matrix may miss several entries. The user-rating
matrixis defined as T  R mn , and to solve the entry miss
problem, a weight matrix W  R mn is defined as well. In
this weight matrix, Wij is equal to 1 when the corresponding
entry is not missing, otherwise Wij is equal to 0. Afterwards,

81
 IV. EXPERIMENT recommendation lists with different lengths ranging from 10
to 100 at intervals of 10.
In this section, we conduct several experiments to test the
performance of our E-commerce recommendation algorithm.
The Movielens[9] is chosen as the dataset, which were
collected by the GroupLens Research Project at the
University of Minnesota. MovieLens datasets consist of two
sub-dataset: 1) 100k and 2) 1M. Moreover, the rating values
are ranged from 1 (Bad) to 5 (Excellent) of users for a
collection of movies in the system. Particularly, the data are
collected through the MovieLens web site
(movielens.umn.edu)
In Movielens dataset, the Movielens rating data are made
up of 100,000 ratings of 1682 movies by 943 users which
are marked from 1 to 5. Before the conducting the
experiment, data pre-process should be implemented in
advance. Five hundred users are chosen who have more Figure 1. NMAE for different length of recommendation list
than 30 ratings on the movies. Moreover, we exploit the
movie class to generate five copies of auxiliary data and
target data. Particularly, for each copy of target data, sixty
percent ratings are regarded as training data, and others are
used for testing.
Afterwards, four metrics are used in the experiment,
which are NAME, Precision, Recall and F1 value. NAME
refers to the normalized mean absolute error, that is one of
the most common approach to evaluate precision of the
recommendation algorithm thourgh comparing users ratings
with the numerical prediction values. Metric NAME is
defined as follows.

N
 rjl  p jl

lTe j

j 1 Te j
NMAE  (5) Figure 2. Precision for different length of recommendation list
N  (rmax  rmin )

In the above equation, N means number of users, and

the maximum and minimum values in the rating range are
represented as rmax and rmin respectively. rjl denotes the
original rating of the product l by user j , and p jl refers to
the predicted rating of product l for the user j
Two other product recommendation schemes are used to
make comparision with our proposed algorithm, that is,
Ratings CF and Tags-CF. The “Ratings CF” approach
compute cosine similarity by ratings which users utilized to
search for neighbors, and then integrate with collaborative
filtering for E-commerce recommendation. The “Tags-CF”
approach calculate cosine similarity by tags which users
utilized to look for neighbors, and then combine the
collaborative filtering technology for E-commerce Figure 3. Recall for different length of recommendation list
recommendation as well.
As the length of the recommendation list may influence
the recommendation results, we evaluate the quality of

82
 [3] Tanbeer Syed K., Leung Carson K., Cameron Juan J., Interactive
mining of strong friends from social networks and its applications in
E-commerce, Journal Of Organizational Computing And Electronic
Commerce, 2014, 24(2-3): 157-173
[4] Lai Jung-Yu, Ulhas Khire Rushikesh, Lin Jian-Da, Assessing and
managing e-commerce service convenience, Information Systems
Frontiers, 2014, 16(2): 273-289
[5] Wang Jing, Ke Liangwen, Feature subspace transfer for collaborative
filtering, NEUROCOMPUTING, 2014, 136: 1-6
[6] Lin Szu-Yin, Lai Chin-Hui, Wu Chih-Heng, A trustworthy QoS-
based collaborative filtering approach for web service discovery,
Journal of Systems and Software, 2014, 93: 217-228.
[7] Zhu Tianqing, Ren Yongli, Zhou Wanlei, An effective privacy
preserving algorithm for neighborhood-based collaborative filtering,
Future Generation Computer Systems-the International Journal of
Grid Computing and Escience, 2014, 36: 142-155
Figure 4. F1 value for different length of recommendation list [8] Nilashi Mehrbakhsh, bin Ibrahim Othman, Ithnin Norafida, Hybrid
recommendation approaches for multi-criteria collaborative filtering,
Expert Systems With Applications, 2014, 41(8): 3879-3900
Combining the experimental results in Fig.1-Fig.4, two [9] GroupLens research. (2014). MovieLens. Available at:
<http://grouplens.org/datasets/movielens/>.
conclusions can be drawn that 1) “Tags-CF” obtain the best
performance than other methods, and the proposed
algorithm performs next only to “Tags-CF”. However,
“Tags-CF” can not provide effective recommendation
results when there are few tags supplied by users. 2) For the
Precision, Recall, and F1 value, the proposed algorithm
performs best in the three methods, and “Tags-CF” is the
worst approach, because target users have identified few
neighbors by tags. Hence, we can known that the proposed
algorithm is intelligent and effective for product
recommendation for E-commerce platform according to
users’ preference.

V. CONCLUSIONS

In this paper, we propose an intelligent E-commerce

recommendation algorithm based onh collaborative filtering
algorithm. First of all, user interest model is deisgned to
describe users’ purchase intention. The main assumption
lies in that users with similar interest may want to buy the
same products. Afterwards, using the E-commerce
recommendation algorithm, products with highest scores
can be recommendated to the target user according to his
preference.

REFERENCES

[1] Hackl Franz, Kummer Michael E., Winter-Ebmer Rudolf, Market

structure and market performance in E-commerce, European
Economic Review, 2014, 68: 199-218
[2] Maity Moutusy, Dass Mayukh, Consumer decision-making across
modern and traditional channels: E-commerce, m-commerce, in-store,
Decision Support Systems, 2014, 61: 34-46

83